chennana1028/gwp05-openarm-sbint-step20000

GigaWorld-Policy-0.5 · SBInt OpenArm (step 20000)

Fine-tuned GigaWorld-Policy-0.5 MoT transformer on SBInt openarm002 (873 episodes, 16-dim absolute joints, left-first packing).

Contents

File	Notes
config.json	CasualWorldActionTransformer_MoT, in/out_action_channels=16
diffusion_pytorch_model.bin	EMA weights from train job 4306868 step 20000
norm_stats_sbint_abs.json	Matching absolute-joint quantile norm stats

Training snapshot

• Job: Slurm 4306868 on 002-partition-RAD, 2×8 H100
• Steps: 20000 · ckpt every 1000 · eff batch 128
• Data: /lustre/.../gwp_sbint/sbint_openarm_v30 (LeRobot v3)
• Base: HF GWP-0.5 + Wan Diffusers VAE via lingbot_va_base
• Code: https://github.com/zhujohn9604/giga-world-policy

Load

from world_action_model.models import CasualWorldActionTransformer_MoT
import torch
model = CasualWorldActionTransformer_MoT.from_pretrained(
    "chennana1028/gwp05-openarm-sbint-step20000",
    torch_dtype=torch.bfloat16,
)

Open-loop helper in the repo: scripts/open_loop_sbint_ep.py.